Chenghao Liu, Zhizheng Zhuo, Liying Qu, Ying Jin, Tiantian Hua, Jun Xu, Guirong Tan, Yuna Li, Yunyun Duan, Tingting Wang, Zaiqiang zhang, Yanling zhang, Rui Chen, Pinnan Yu, Peixin Zhang, Yulu Shi, Jianguo Zhang, Decai Tian, Runzhi Li, Xinghu Zhang, Fudong Shi, Yanli Wang, Jiwei Jiang, Aaron Carass, Yaou Liu, Chuyang Ye. "DeepWMH: a deep learning tool for accurate white matter hyperintensity segmentation without requiring manual annotations for training". Science Bulletin, Jan 2024.

Supplementary materials: https://drive.google.com/file/d/13Kyk0v19kRnJ40JC22-p_YK2urnzZcvh/view?usp=sharing

• About DeepWMH
• Python Requirement
• Quick start: how to use our pretrained model (only for Linux-based systems)
• Advanced: how to train a model using data of my own? (only for Linux-based systems)

• If you have any questions or find any bugs in the code, please open an issue or create a pull request!

DeepWMH is an annotation-free white matter hyperintensities (WMH) segmentation tool based on deep learning, designed for accurately segmenting WMH lesions using T2FLAIR images without labeled data. An overview of the whole processing pipeline is shown below.

The figure below shows a more detailed version of the processing pipeline. Please refer to the supplementary materials for more information.

Before installation, you may need to update your Python version if it is lower than "3.7.1". Also, this tool is based on Python 3, Python 2 is deprecated and should no longer be used anymore.

The fastest way of applying our tool to your research is by using our pre-trained model directly. To use our pre-trained model for inference, please follow the steps below:

1. Update your Python environment. Then, create a new virtual environment using the following commands:

   pip install -U pip                         # update pip
   pip install -U setuptools                  # update setuptools
   pip install wheel                          # install wheel
   python -m pip install --user virtualenv    # install virtualenv
   python -m venv <path_to_your_virtual_env>  # create a virtual environment under <path_to_your_virtual_env>

   for more detailed information about how to create a Python 3 virtual environment, please refer to this link.

2. Activate the virtual environment you just created:

   source <path_to_your_virtual_env>/bin/activate

   NOTE: the virtual environment should ALWAYS be activated during the following steps.

3. Install PyTorch in your virtual environment. See https://pytorch.org/ for more info.

4. Download nnU-Net source code from "https://github.com/lchdl/nnUNet_for_DeepWMH".

   PLEASE download the forked version above, DO NOT download from https://github.com/MIC-DKFZ/nnUNet as the forked version has made some necessary changes.

   Unzip the code, "cd" into the directory where setup.py is located, then execute the following command:

   pip install -e .

   This will install customized nnU-Net and all its dependencies into your environment.

   Please install PyTorch BEFORE nnU-Net as suggested in https://github.com/MIC-DKFZ/nnUNet, and make sure your CUDA/cuDNN version is compatible with your GPU driver version.

5. Download DeepWMH from "https://github.com/lchdl/DeepWMH". Unzip the code, "cd" into the directory where setup.py is located, then execute the following command:

   pip install -e .

6. Download and unzip ROBEX (Robust Brain Extraction) from "https://www.nitrc.org/projects/robex",

   

   then add:

   export ROBEX_DIR="/path/to/your/unzipped/ROBEX/dir/"

   in your ~/.bashrc, make sure "runROBEX.sh" is in this directory, then

   source ~/.bashrc

   to update the change.

7. (Optional) compile & install ANTs toolkit from "https://github.com/ANTsX/ANTs", or download the pre-compiled binaries here. This is mainly for intensity correction. You need to add ANTs binaries to "PATH" in your ~/.bashrc before using them.

   NOTE: You can also skip this step if you don't want to install it. However, the segmentation performance can be seriously affected if the image is corrupted by strong intensity bias due to magnetic field inhomogeneity. For optimal performance I strongly recommend you to install ANTs.

   Verify your install: to see whether ANTs is installed correctly in your system, after the installation you need to type in

   antsRegistration --version
   

   and

   N4BiasFieldCorrection --version
   

   in your console. It should produce output such as:

   ANTs Version: 3.0.0.0.dev13-ga16cc
   Compiled: Jan 22 2019 00:23:29
   

   Then test if antsApplyTransforms can work:

   antsApplyTransforms
   

   if no error shows, then ANTs is successfully installed.

8. (Optional) verify your install:

   1. activate your virtual environment
   2. enter Python by typing and running:

   python

   3. then, enter & run the following script line by line:

   from deepwmh.main.integrity_check import check_system_integrity
   check_system_integrity(verbose=True, ignore_ANTs=True, ignore_FreeSurfer=True, ignore_FSL=True)

   4. if something is missing, you will see some error messages popping up. Some tips about how to fix the errors are also given. You can follow the tips to fix those problems and repeat Step 8 to verify your install until no error occurs.

9. After installation, run

   DeepWMH_predict -h

   if no error occurs, then the installation is complete! Now you are ready to use our pretrained model for segmentation.

10. Download our pre-trained model (~200 MB) from

    1. "https://drive.google.com/drive/folders/1CDJkY5F95sW638UGjohWDqXvPtBTI1w3?usp=share_link" or
    2. "https://pan.baidu.com/s/1j7aESa4NEcu95gsHLR9BqQ?pwd=yr3o"

    Then use

    DeepWMH_install -m <tar_gz_file> -o <model_install_dir>

    to install model (as indicated by <tar_gz_file>) to a specific location (as indicated by <model_install_dir>).

11. Using pre-trained model to segment WMH lesions from FLAIR images with the following command:

    DeepWMH_predict -i <input_images> -n <subject_names> -m <model_install_dir> -o <output_folder> -g <gpu_id>

    if you don't have ANTs toolkit installed in your machine (see Step 7 for how to install ANTs), you need to add "--skip-bfc" to the end of the command, such as:

    DeepWMH_predict -i <...> -n <...> -m <...> -o <...> -g <...> --skip-bfc

    the segmentation performance can be seriously affected if the image is corrupted by strong intensity bias due to magnetic field inhomogeneity

    NOTE: you can specify multiple input images. The following command gives a complete example of this:

    DeepWMH_predict \
    -i /path/to/FLAIR_1.nii.gz /path/to/FLAIR_2.nii.gz /path/to/FLAIR_3.nii.gz \ # specify three FLAIR images
    -n subject_1               subject_2               subject_3 \               # then give three subject names
    -m /path/to/your/model/dir/ \                                                # model installation directory
    -o /path/to/your/output/dir/ \                                               # output directory
    -g 0                                                                         # gpu index

1. Follow the Steps 1--7 in Quick start section. Note that if you want to train a custom model, you must download and compile ANTs toolkit, Step 7 in Quick start section is no longer optional.

2. Download and install FreeSurfer. Note that you may also need to install "csh" and "tcsh" shell by running

   sudo apt-get install csh tcsh

   after the installation.

   A license key (link) is also needed before using FreeSurfer.

3. Download and install FSL (FMRIB Software Library).

   How to install: FSL is installed using the fsl_installer.py downloaded from here. You need to register your personal information to the FSL site before download. After download you need to run the installer script and wait for the installation to finish.

   Verify your install: when the installation finished, type in

   bet -h
   

   in your console. If no error occurs then everything is OK! :)

4. (Optional) verify your install:

   1. activate your virtual environment
   2. enter Python by typing and running:

   python

   3. then, enter & run the following script line by line:

   from deepwmh.main.integrity_check import check_system_integrity
   check_system_integrity()

   4. if something is missing, you will see some error messages popping up. Some tips about how to fix the errors are also given. You can follow the tips to fix those problems and repeat Step 4 to verify your install until no error occurs (as shown below).

      

5. Here we provided two examples of using a public dataset (OASIS-3) to train a model from scratch, see

   experiments/010_OASIS3/run_Siemens_Biograph_mMR.py
   experiments/010_OASIS3/run_Siemens_TrioTim.py
   

   you can also run these two examples if you organized the dataset structure correctly:

   cd .../experiments/010_OASIS3/
   python run_Siemens_Biograph_mMR.py

   cd .../experiments/010_OASIS3/
   python run_Siemens_TrioTim.py